Volta 16k

Discussion in 'Videocards - NVIDIA GeForce' started by DreadZilla101, Sep 19, 2017.

Tags:
?

Am I jumpung the gun or was this interesting?

  1. Jumping the gun

    14.3%
  2. Interesting

    42.9%
  3. You're just a goof lolz

    42.9%
  1. DreadZilla101

    DreadZilla101 Member

    Messages:
    21
    Likes Received:
    0
    GPU:
    GTX 660 Ti/2048 MB
    Hello everyone,

    This is my first thread and the calculations might be based more on conjecture than anything else, but there is a slight chance that everything I theorize is right so I decided to make this thread. There is much opinion involved, I hope that everything checks out.

    This post assumes that Nvidia follows a pattern where their first Tesla GPU is only a fraction of the capabilities of their GPUs designed for more graphically demanding roles.The only example of Volta architecture we have now is the Tesla V100, clocking in at 15 Tflops FP32. The Pascal Tesla came in at 10.6 FP32 and the final release the Titan Xp was 12.15 FP32, a 12.8% (or .127572016) increase. So the projected maximum for Volta should be 17 Tflops ([15*.127572016]+15=16.91358024, rounded to 17).

    I theorize that based on the scaling of the 1080 TI in Dawn of War 3 (a demanding game designed for PC), based off benchmarks at high settings at both 4k and 1080p provided at notebookcheck.com), that quadrupling resolution causes a 66% (.65510204) performance hit (it dropped from 98 average frames per second to 33.8 at only High settings in 4k, not even Ultra).

    The GTX 1080 had 25.7 frames per second at High settings in 4k, a 24% (.23964497) decrease when compared to the 1080 TI. The 1080 is at 9 Tflops FP32 and the 1080 TI is at 11.34, a difference of 2.34 Tflops. Ignoring all other relevant factors (like VRAM speed and quantity), if 2.34 Tflops equals a 24% performance increase in Dawn of War 3 (my chosen point of reference), then a 5 Tflop increase in FP32 will result in a 51% increase in average frames per second ([5/2.34]*.23964497=.512061901).

    All of this means that at high settings in 4k a single, non-overclocked, 17 Tflop Volta GPU could raise average frames per second (in Dawn of War 2 at 4k, high settings) to 51 ([33.8*.512061901]+33.8=51.10769225).

    If Volta GPUs scale well it would take Quad SLI scaling at 95% to reach 162 frames per second (33.8+[33.8*.95]*4=162.24) in Dawn of War 3 at 4k, high settings. At 8k high settings it could bring 4 non -overclocked high level Volta GPUs (probably Titans) down to 56 frames per second (162.24-[162.24*.65510204] =55.95624503).

    At 16k the Quad SLI setup (with non-overclocked cards) would theoretically manage only 19 frames per second (55.95624503-[55.95624503*.65510204]=19.29919476) in Dawn of War 3 16k, at high settings.

    4 Quadro P5000s (each at 8.9 Tflops FP32), ran Rise of the Tombraider at 1 fps in 16k couldn't have been running over 3 fps in Crysis 3 16k (though they did manage to run Minecraft at an apparent average fps of 50, just flying around slowly in creative mode), during the LinusTechTips experiment.


    (I apologize for the implied curse word in the title.)

    Instead of SLI, Linus used Nvidia Quadrosync which he wrongly assumed would overcome the VRAM constraint (limited to the maximum of one card instead of the total of all 4 [hopefully Nvidia will fix this]) found in SLI setups. While it worked fine for Half-Life 2 and Minecraft at 16k, the 16GB of VRAM of the P5000 was not enough for 16k (current Quadro P6000s have 24GBs of GDDR5X) in Rise of the Tombraider and other graphically intensive titles (at what appeared to be low settings).

    Linus believed that since the P5000s were running at 66% and 100% of VRAM (16GB) was in use, he could have gained a possible 30+% in performance if the setup had more VRAM. The increase, he said, would have brought performance to "1.3" frames per second. This suggests that when gaming at super high resolutions 24 GB is the minimum amount of VRAM required.

    Linus's setup seemed to have worked better than traditional SLI (without Venturi), though the scaling of the Linus setup was not fully disclosed. There should be a Linus/Venturi collaboration build, we should all start trying to make that happen. That would be insanely awesome.

    If the high level Volta cards have at least 24 GB of VRAM, Volta Quadro V6000s may be needed (The P6000 is almost equal to the Titan Xp, 12 Tflops for the P6000 and 12.15 Tflops for the Titan Xp), could run a graphically intensive game (like Dawn of War 3) in Venturi-SLI far better than the Linus original build did (at a theoretical average of 19 frames per second).

    Also, I believe SLI will work better with overclocking (water cooling would be needed) than Quadrosync, so we need a Quadro-Venturi-SLI vs Quadro-Quadrosync.

    If we can learn from the past then the V6000 and Titan Xv could be held off until late 2018 for the V6000 and sometime in 2019 for the improved Titan Xv.

    To conclude, we need a Venturi-Linus team up, even future cards will struggle with 16k (I'm Cpt. Obvious) and it is crazy how natural it is to overlook 8k and 4k when you know that real Quad SLI really exists. 16k should be the last resolution for a while though, until some major technological breakthrough. Though at our current rate of progression, there will be setups running 16k at 60 fps by 2022.
     
    Last edited: Sep 20, 2017
  2. Agent-A01

    Agent-A01 Ancient Guru

    Messages:
    10,376
    Likes Received:
    260
    GPU:
    1080Ti H20
    One thing to note is that stock boost rated tflops is 11.2 at 1582mhz.

    A 2100mhz 1080ti is about 15.3~ tflops.
    An average-ish OC of 1974mhz is 13.9 tflops.

    Every 177mhz gain is about 1.3tflops.

    Lets say volta clocks similarly and also has a similar gain in tflops(this is just pure speculation but we can assume it will clock higher with tensor cores cut out)

    At 1455mhz V100 has 15tflops.
    A 1080Ti at 1455mhz will be roughly 10.3 tflops.

    By using Ti baseline gain from OC, a ~2ghz V100 will 19~ tflops.
    That's pretty impressive but I expect the scale in tflops to be bigger on volta considering it has 42% more cores(comparing V100 and Ti)

    If tflops scaled 42% better than a 2ghz V100 would have 20.5 Tflops.

    Anyways, i really hope it is a significant improvement.
    4k 120hz is around the corner and even two 1080tis will not be enough.
     
    DreadZilla101 and airbud7 like this.
  3. DreadZilla101

    DreadZilla101 Member

    Messages:
    21
    Likes Received:
    0
    GPU:
    GTX 660 Ti/2048 MB
    Wow, thanks for your insight. It is interesting to see an overclock expressed in terms of teraflop gain (also, lol I called the V100 Tesla a V1000 [changing that now]).

    I tried to leave the overclockability out of my assessment because heat becomes such a concern that without adequate watercooling anything significant is unlikely, sometimes it feels like a misrepresentation of the hardware (but it would have been much more efficient to include the overclocking scale, as I was trying to speculate with enthusiasts in mind).

    Also, in your example you used the 1080ti (at 11 GB of GDDR5X, isn't it restricted when compared to the Titan Xp [12GB GDDR5X] and Quadro P6000 [24GB GDDR5X]?), was it to represent the consumer oriented GPUs?

    As you know better than I do, the Titan Xp has the highest baseclock of any in the pascal lineup so I used it to represent the highest teraflop count in my assumption (but the P6000 is more efficient for higher resolutions, with an aforementioned insane 24GB of GDDR5X VRAM).

    I hope the Volta architecture scales to 20.5 teraflops (we might have overclocked to 16k at 60hz around 2020, with Quad SLI and around 24 to 32GB of GDDR7X VRAM). And I hope that Nvidia releases some kind of SLI RAID0 to get past the VRAM constraint.

    Also I wanted to ask if you've watched any of ThirtyIR's videos on YouTube (Venturi's configuration is much more stable)? What do you think of ThirtyIR's Quad SLI stability? And do you believe that Quad SLI will have a more major role in the future because of Venturi's solutions?

    Also, thanks again for your response, you've helped provide a lot of context.
     
    Last edited: Sep 20, 2017
  4. sneipen

    sneipen Member

    Messages:
    28
    Likes Received:
    5
    GPU:
    Nvidia 1080 2ghz
    Looking forward to the day when 4k performance is like 1440p today.
    My oc'ed gtx1080 does the job ok but still i would like a bit more grunt, but i feel selling my 1080 to buy a ti isnt worth the hassle . So, waiting for volta or amd's navi and im hoping that generation will have what it takes, maybe also rund 4k+ vr..
    Would be great if AMD next gen. can create some competition.
    Gpu marked has been #¤%&/ for to long...
     
    DreadZilla101 likes this.

  5. DreadZilla101

    DreadZilla101 Member

    Messages:
    21
    Likes Received:
    0
    GPU:
    GTX 660 Ti/2048 MB
    Without an overclock the difference between the GTX1080 (9 Tflops) and 1080ti (11.34 Tflops), and even the 1080ti is not the most powerful representative of pascal architecture, is 8.1 fps in a demanding (made for PC, so it scales well) game like Dawn of War 3. The 1080ti only managed 33.8 frames per second, in Dawn of War 3 in 4k, at stock settings.


    For a single card to reach 60 frames per second in such a game (ignoring the importance of the CPU) it would take a theoretical 19.14 Tflops ([7.8/2.4]*.23964497=.778846153).

    A +78% ([.778846153+1]*33.8=60.12499997) increase in teraflops, bringing the average frame rate to 60. So with two 1080s you'd reach 60 fps at 4k with no problem (just a slight overclock).


    But when you say 1440P one could assume you meant 120 frames per second. For a single card to run such a game in 4k at 120fps it could take 36.84 Tflops ([25.5/2.4]*.23964497=2.546227806).

    A +255% ([2.546227806+1]*33.8=120.1999999) increase in teraflops. If we get an overclock to 12.3 Tflops for each GPU, a +27% increase (-[9/12.3]+1=.268292682), it would still take 3 (36.84/12.=2.995121951) GTX1080s in SLI with perfect scaling to reach 120fps at 4k in such a game (likely by 2020 a single Titan of that era overclocked to the extreme could pull of 4k 120hz).

    I'm currently running a 660ti on a system that was fine 5 years ago, I got it refurbished system from Cyberpowerpc, through Amazon. It was far below the market price of all components involved (due largely to the powerful CPU). I upgraded the PSU and GPU (from a GTX650), but I'd be struggling with 1080p in these new titles (without an upgrade) if I still played PC games.

    So, I agree with you to the extreme, I'm waiting to see what Volta can do (I hope it's a good surprise).
     
  6. Hammie

    Hammie Banned

    Messages:
    707
    Likes Received:
    8
    GPU:
    760 SMG 28"
    That is so ugly, little square monitors ,,,, wanna see the pic on the borders of the monitor. What a pos imo. gl
     
    DreadZilla101 likes this.
  7. Xanvast

    Xanvast Master Guru

    Messages:
    289
    Likes Received:
    3
    GPU:
    Titan X P 3 Way SLI
    The drop in performance compared to the resolution increase is linear with cpu and other bottlenecks excluded.
     
    DreadZilla101 likes this.
  8. DreadZilla101

    DreadZilla101 Member

    Messages:
    21
    Likes Received:
    0
    GPU:
    GTX 660 Ti/2048 MB
    Yeah, I don't really understand multi monitor setups which use multiple monitors as one screen. I'd rather one giant, wide, curved screen (he could have used multiple 4k projectors).

    16k is not looking possible on a small scale right now (I've never seen a small 16k prototype). But I think future me would prefer a 16k 40 inch screen. By then Nvidia could use a kind of RAID0 so SLI cards share memory more effectively (as Linus's experiment proved), eliminating the VRAM bottleneck (4 P6000s of today would have a shared 96GB of GDDR5X VRAM).

    Nvidia also has to drastically increase SLI bandwidth between cards, making them basically form Voltron (or a Megazord), working as one unit and scaling perfectly across all cards involved (Ultimate SLI). And officially recommend more than 1 CPU (the Venturi way) to manage the insane amount of resources which future enthusiasts will command.

    I think Venturi has proven that the single CPU approach is becoming obsolete at the enthusiast level, very interesting to see how things develop.
     
  9. DreadZilla101

    DreadZilla101 Member

    Messages:
    21
    Likes Received:
    0
    GPU:
    GTX 660 Ti/2048 MB
    Does anything you've seen support that quadrupling resolution reduces performance by ~66% (that was my guess)? Or do you believe it's scaling differently?
     
  10. Xanvast

    Xanvast Master Guru

    Messages:
    289
    Likes Received:
    3
    GPU:
    Titan X P 3 Way SLI
    Quadrupling the resolution more or less divides the performance by 4 / reduces it by 75%
     
    DreadZilla101 likes this.

  11. DreadZilla101

    DreadZilla101 Member

    Messages:
    21
    Likes Received:
    0
    GPU:
    GTX 660 Ti/2048 MB
    That does make sense mathematically (in a theoretical sense), but what I see in benchmarks is strangely contradictory. What you're saying is that the result of quadrupling should be the original frame rate multiplied by .25 (scary if it was like this, but why is it not?).

    How does average frame rate not scale by your logical assumption? It seems like under lower resolutions 100% of the GPU's available resources might not be in use, possibly the result of vsync. If this is correct then higher resolutions often stress GPU's to use much more energy (which seems like I'm stating the obvious as I type it), resulting in increased performance

    This increase somehow brings performance up by almost 10%. Performance = >25%,<35%, I wonder how exactly it happens (all I have are really hypotheses). I believe that both the resources available to the GPU and it's efficiency get a bump at higher resolutions.
     
    Last edited: Sep 22, 2017
  12. DreadZilla101

    DreadZilla101 Member

    Messages:
    21
    Likes Received:
    0
    GPU:
    GTX 660 Ti/2048 MB
    Maybe Nvidia could design cards in SLI from the beginning and sell them as SLI unit blocks. With each interconnected using the fastest transmission between GPUs possible with current technology.

    Everyone who wants to buy GPUs separately could still do that but suffer a slight performance decrease, or just save up your money to buy the SLI you want (like a Duel, Tri or Quad SLI block).

    Each block would be double spaced to fit in any motherboard, and register as one GPU in the system with each (of 2 to 4 components) manageable like a single card (just fully synchronized).

    So if developers want to manage resources between each card they can, though all VRAM would be shared across cards in a RAID0-like configuration.
     
  13. sneipen

    sneipen Member

    Messages:
    28
    Likes Received:
    5
    GPU:
    Nvidia 1080 2ghz
    Thata what im thinking where amd is going, seeing what thredripper offers. It seems like they found out how to scale cpu. I belive this is the roadmap. Take this ability to gpus. And if they manage to make 2+ gpus scale effective they can really push nvidia even tho one gpu performene perfome worse.
     
    DreadZilla101 likes this.
  14. DreadZilla101

    DreadZilla101 Member

    Messages:
    21
    Likes Received:
    0
    GPU:
    GTX 660 Ti/2048 MB
    I agree, I think that AMD is a major factor in making Nvidia do what they need to and make a better SLI.
     
  15. venturi

    venturi Member Guru

    Messages:
    150
    Likes Received:
    18
    GPU:
    4x Titan V
    Again,

    A lack of competition from and for the past 4 years has created the urgent situation.

    Also bear in mind that nvidia does not want to have to justify the price of their high end ai/research division products if the regular joe can achieve 80% of that performance in a regular prosumer workstation using their own products.


    As far as the 16k video/Linus station....

    ... I could have SLIed that :)

    Bit of a waste not to
     
    The1, DreadZilla101 and Xanvast like this.

  16. DreadZilla101

    DreadZilla101 Member

    Messages:
    21
    Likes Received:
    0
    GPU:
    GTX 660 Ti/2048 MB
    It is amazing how situations like this can be formed, I hope AMD forces Nvidia to change their ways.

    It would be great to see you maximize what Linus did and show everyone what 4-way SLI can really do.
     
  17. HeavyHemi

    HeavyHemi Ancient Guru

    Messages:
    5,320
    Likes Received:
    34
    GPU:
    GTX1080Ti
    It's amazing how so many threads turn into 'It 's AMD's fault Nvidia has not....' based on nothing but their expectations of performance increases that are simply not in the pipeline or technologically feasible. What technology or advancement has Nvidia not done BECAUSE OF, AMD? How about a list?
     
    DreadZilla101 likes this.
  18. DreadZilla101

    DreadZilla101 Member

    Messages:
    21
    Likes Received:
    0
    GPU:
    GTX 660 Ti/2048 MB
    AMD is not the only factor in the adoption of new GPU technology, but you know that if AMD starts making dual-socket motherboards mainstream (and unified VRAM in Crossfire) in a last ditch effort to win the war, Nvidia will have to respond.
     
  19. venturi

    venturi Member Guru

    Messages:
    150
    Likes Received:
    18
    GPU:
    4x Titan V
    You are correct.

    Its not AMD's fault.

    Nvidia was just more formidable and made better decisions. Intel and Nvidia did great, AMD did well. As far as market / sector percentage - Intel and Nvidia do better, AMD does a nice product, Intel and Nvidia do a better one when price is not a deciding factor.

    In the world of corporate strategies, Nvidia is pacing itself and AMD is running as fast as it can.
    Example, nvidia release a driver update on AMD's release week that boost the titan xp 300% in some applications...while AMD has to consume a lot more power and greater cooling solutions just to match with the non the top of the line product from nvidia.

    Its not AMD's fault. I like AMD. If you go back far enough many of my dual and quad processor workstations were AMD. I cheer for them and want them to win, completely. I have built so many uber workstations stations. Every time trying desperately to use AMD.

    With that said, nvidia unfortunately in several respects has a better product and a release path that AMD wont be able to get ahead of it. The various bulldozer and sidelined chips of the past 7 years have not helped AMD's cause. So now AMD is a plucky company trying hard and I want them to win. They will have to out-innovate the others. I'm hoping they will, but realistically its a longshot for them.

    AMD's market pattern is that when they can't beat nvidia products -they retreat to a performance/price per model approach. I respect that, but by doing that they skew their perception to 2nd place.

    AMD needs their best product to beat Nvidia's best product, it needs to a be a slaughter, something that will cause tremors at nvidia (what I want)


    So you are correct, its not their fault and they try hard. I realize I'll get flamed and trolled but the reality is as such: AMD has not been significant competition to nvidia, -not significant enough for nvidia to not have pulled the shenanigans they do:

    1.telemetry in drivers. and apps
    2. performance gimping to favor other product vertical sales
    3. killing SLI
    4. sky high price points


    seriously:

    I want an AMD top of the line pro-sumer card to beat mercilessly nvidia's top of the line pro-sumer card without using 50% more power and require uber water-cooling. Its what I want, and I want it to be a slaughter.

    Well, I'd buy right as many top of the line AMD cards as I could fit on a single (dua/quad) motherboard if they could do that and buid the ULTIMATE workstation (insert Dr. Evil cruel chuckle)

    So make no mistake, I want them to win.

    I want to go on to newegg or amazon and buy an AMD card (s) that beats the titan XP by as little as 10% on average for my games or apps.
    I want that choice. Where is that choice?

    I want to build a 4 or 6 way mgpu dual/quad socket threadripper workstation using only AMD chips, board and cards - that will outperform my current build. May I please have that option?


    Can we all at least be honest and observant?

    nvidia needs significant competition. Intel needs significant competition. Is AMD the company that will beat their top of the line products? Soon?


    I think I have the most powerful Quad SLI rig, but would have gladly built my rig with AMD products if I had the choice (in terms of performance)
    I want that choice!
     
    Last edited: Sep 26, 2017
    The1, DreadZilla101 and Xanvast like this.
  20. Agent-A01

    Agent-A01 Ancient Guru

    Messages:
    10,376
    Likes Received:
    260
    GPU:
    1080Ti H20
    I'm just going to put out the fact that you are a fanboy, which i don't care about in the sense, but those 4 reasons you listed as 'sh1t nvidia pulls because of amd's lack of significance' is just wrong.

    1. Telemetry is not like windows, they gather application data and collect crash logs for bug fixes.
    2. They don't purposely gimp performance. If you are referring to Titan XP boost; that is a surprise as that card is not marketed as a quadro card yet got quadro profiles for certain applications due to that $1000+ vega pro AMD press release bench comparison.
    Neither of those two cards are marketed for pure gamers.
    3. The only thing 'killed' was quad SLI, even though that's still possible to use. Considering most newer games are up to game dev to support multigpu, this is a moot point.
    4. Prices are actually cheaper than they were 10 years ago for the performance you get. Pascal is by far the better deal right now, AMD prices are sky high.
    Vega56 is more expensive than a 1080Ti in some places. Vega 64 is more expensive than the Ti period while not even comparing to the Ti.
     

Share This Page